Classifying Tumour & Immune Cells in H&E Images
Motivation - Breast Cancer
Research Direction
Goal: Automatically identify tumour and immune cells in H&E tissue images.
Challenge: Find model fit for usage in clinical settings.
Approach: Evaluate & compare 3 different classification models.
Workflow
- Image Data Preparation & Processing
- Model Parameters & Training
- Evaluation & Selection
- Clinical Implementation
Candidate Models
- Random Forest Classification
- Convolutional Neural Network
- Support Vector Machine
1 - Random Forest
- Feature Extraction (HOG)
- Dimensionality Reduction using PCA
- Random Forest Classification
1 - Random Forest
1 - Random Forest
2 - Convolutional Neural Network
- Automatic Feature Learning
- CNN Architecture Design
- Training Process
2 - Convolutional Neural Network
3 - Support Vector Machine
- Convert to Grayscale
- Feature Extraction
- SVM Classification
- Optimizationm using CV
3 - Support Vector Machine
Evaluating Candidate Models
Preferred Model - CNN
Outperformed Other Models
High Sensitivity (95%): High detection of tumour cells.
End-to-End Learning - No need to extract or select features
Automatic Feature Extraction - Finds more subtle image patterns hierarchically
Scalability - Can handle larger datasets with more cell types
Limitations
Large data requirements
CNN “Black box” blurs interpretability
More computationally expensive than RF and SVM
Prone to over-fitting
Takeaway
Clinical Implications: CNN had the best sensitivity to tumour cells.
Professional Interpretation: Positive test results must be carefully examined by pathologists.
Future Direction: Look at classifying other cell types, examining other tissue and using different imaging systems
References
ChatGPT (OpenAI, 2023) was used for assistance in preparing this presentation template, debugging and searching for additional packages for models.